TraceLingo: Trace representation and learning for performance issue diagnosis in cloud services

Large-scale cloud systems such as Microsoft Azure, Google Cloud and Amazon AWS provide a wide variety of online services which serve millions of customers around the world. Adverse service behaviors and latencies can have huge performance impact and affect user satisfaction. Besides monitoring system KPI metrics and log, trace data is of great value for analyzing the system performance status, detecting anomalous workstreams and localizing the performance bottleneck. However, existing work usually represent the trace as a sequence of events with execution time information, which ignores the runtime context and graph structure of the trace. In this paper, we propose a trace representation and learning model, TraceLingo, which adopts a tree-based RNN model to capture the dependency between spans in various traces for automatic and effective performance diagnosis.